OpenAI shut down its preparedness team at the end of last month, according to the Financial Times — eliminating the internal group whose job was to evaluate whether its models posed serious risks and build ways to contain them.
Key takeaways
OpenAI disbanded its preparedness team — responsible for catastrophic risk assessment and mitigation — at the end of last month, per the Financial Times.
The team's duties are being redistributed to other internal groups, though the exact structure of that handoff has not been publicly detailed.
The preparedness team's mandate included scenarios like a model autonomously hacking external systems — risks directly tied to the increasingly agentic AI tools creators use daily.
The move comes as OpenAI faces ongoing scrutiny over its safety culture following high-profile departures of safety-focused researchers.
No changes to current model behavior or API access have been announced as a result of the restructuring.
OpenAI's preparedness team, which assessed catastrophic model risks, was disbanded at the end of last month according to the Financial Times.
The Verge reports that responsibility for the work is being folded into other teams, though OpenAI has not publicly detailed which teams absorb which functions. That opacity matters: the preparedness team was specifically chartered to stress-test frontier models against worst-case scenarios — including the possibility of a model autonomously compromising external systems — before those models shipped.
What the preparedness team actually did
The preparedness team sat between research and deployment, running structured evaluations on OpenAI's most capable models to identify catastrophic failure modes before public release. Think of it as a dedicated red-team unit with a mandate to ask whether a given model could, say, assist in creating biological weapons or take autonomous harmful actions on the internet. Those aren't hypothetical framings — they were the team's actual evaluation categories.
For AI creators, this is a few steps removed from the prompt you typed this morning. But the models you render with, upscale through, or use to batch-generate character references are the same frontier systems that preparedness was evaluating. When that evaluation function gets diffused across teams without a clear owner, the question of who catches a dangerous capability before it ships becomes genuinely harder to answer.
The disbanding also lands in a fraught moment. OpenAI has seen a string of safety-focused researchers exit over the past year, and the company converted from a nonprofit to a for-profit structure earlier this year — a transition that drew criticism from former employees who argued it weakened accountability. Removing a dedicated safety evaluation team, even if the work nominally continues elsewhere, fits a pattern critics have been tracking closely.
The agentic angle creators should watch
The preparedness team's scope included agentic risk — scenarios where a model takes multi-step autonomous actions, including potentially harmful ones. That's not an abstract concern anymore. Agentic workflows are increasingly part of how creators operate: models that browse references, execute code, manage files, or chain together tasks without step-by-step human approval. The more autonomous the pipeline, the more the underlying safety evaluation of that model matters.
If you're building agentic setups on top of OpenAI's APIs — chaining GPT-4o calls to manage assets, generate variations, or interact with external services — the question of who inside OpenAI is rigorously evaluating those capabilities for failure modes is now less clear than it was a month ago. That doesn't mean the models are suddenly unsafe, but it does mean the institutional structure for catching problems before they reach you has changed.
For creators who've been watching the broader AI safety picture, recent incidents elsewhere in the industry illustrate why dedicated evaluation matters — the Grok image-safety failure being one stark example of what happens when guardrail evaluation misses a real-world attack vector.
OpenAI hasn't commented publicly on the specifics of how preparedness work will continue. Until it does, the honest read is: the function exists in some form, but the dedicated team that owned it doesn't.
